Loading
Old version — revision 1
This is a fixed snapshot of Human digital twins, saved by Import as part of the initial corpus import. It is not edited and it is not updated; the article may have changed since.
Edit summary: Initial import of content/human-digital-twins.md — the filesystem corpus, unchanged. Not an edit.
Computational models of an individual person's physiology, built from that person's own data and run to predict what an intervention would do before it is attempted.
Human digital twins are computational models of an individual person's physiology, built from that person's own imaging, measurements and sensor data and run to predict what an intervention would do before it is attempted. The idea is imported from aerospace and manufacturing, where a simulation of one specific engine is kept synchronised with the engine itself and used to decide when to service it. Medicine has no whole-body equivalent and is not near one. What exists is a set of organ- and device-specific models, each narrow, each personalised in a few respects and generic in the rest.
Three different objects travel under the name, and the difference between them is most of the subject.
The first is a mechanistic simulation whose geometry is taken from a patient's scan and whose equations come from physics and physiology: blood flow through that person's coronary tree, electrical propagation around the scar in that person's ventricle, the field a stimulation lead produces in that person's brain. The second is a statistical or machine-learned predictor fitted to population data and evaluated on one person's record, closer in kind to a biological-age estimate than to a simulation. The third is a longitudinal data display: a record assembled from Wearable health sensors, Continuous glucose monitoring and the laboratory, rendered next to a picture of a body.
Only the first can be intervened on, which is what the word twin promises. Vendors apply it to all three, and hospital software sold as a digital twin platform is usually the third. Self-tracking products have inherited the vocabulary as well. The distinction worth holding onto is between a model personalised with an individual's own measurements and a model carrying population parameters that merely displays that individual's numbers.
Not an emulationA digital twin is a predictive instrument, judged by whether its outputs match measurements taken from the body it models. Whole brain emulation proposes something categorically different: a simulation detailed enough that it might itself be a bearer of the original's mental life, which is the premise Mind uploading rests on. Treating the two as points on one scale imports the whole of Personal identity and continuity into what is otherwise an engineering question about model error.
A patient-specific model is assembled in four steps, and the third decides whether the object deserves the name.
Geometry comes first. A CT or MRI volume is segmented into a mesh of the relevant anatomy, a step now largely automated and genuinely individual: the mesh is that patient's artery, that patient's infarct. Physics comes second, as constitutive equations selected from the literature — a fluid model for blood, an ionic model for a cardiac cell, a compartmental model for a drug distributing through tissue.
Third is personalisation, meaning estimation of the model's parameters for this person. A cardiac electromechanics model carries tens to hundreds of parameters, far more than the independent measurements obtainable from a living patient. Most parameters are therefore fixed at population values while a small identifiable subset is tuned until the model reproduces something that was actually measured. The commercially deployed case makes the pattern plain: in CT-derived coronary flow analysis the arterial geometry belongs to the patient, but microvascular resistance and resting flow are assigned from scaling relationships between vessel size, myocardial mass and population physiology, because they cannot be measured non-invasively in that patient.
Fourth is the counterfactual run and its uncertainty. The model is re-solved with a stent in place, a lesion ablated, a dose changed. A prediction without an uncertainty interval cannot support a decision, and quantifying that interval is harder than the simulation itself, because it requires knowing how wrong the fixed population parameters might be.
A twin in the engineering sense is also supposed to stay synchronised, ingesting new data and re-estimating as the system changes. Almost no medical model does this. Most are built once from one scan, used for one decision, and then discarded.
The name arrived from aerospace, where pairing a physical vehicle with a virtual model continuously updated from its sensors was proposed as a way to manage airframe fatigue over a lifetime of flights.1
The most widely deployed patient-specific simulation in medicine is decades old and is rarely described as a twin. Radiotherapy planning computes the dose a beam arrangement will deposit throughout that patient's own CT volume, and treatment is prescribed from the model's output.
In cardiology, CT-derived fractional flow reserve estimates how far pressure falls across a coronary narrowing by solving flow equations on the reconstructed artery, sparing some patients a catheter. It is cleared for marketing, reimbursed in several health systems, and was evaluated against invasively measured pressure in patients already scheduled for catheterisation.2 Separately, models of arrhythmia built from contrast-enhanced MRI of a patient's scar have been used to propose ablation targets, the simulation pacing the virtual heart to find where reentrant circuits can be induced.3 That work has been tested in small studies, not in a large randomised trial.
In-silico testing of devices and drugs is further along than in-silico treatment of individuals. A metabolic simulator of type 1 diabetes was accepted as a substitute for animal testing in the preclinical evaluation of closed-loop insulin controllers, and it was used in developing control algorithms that went on to reach clinical use.4 FDA researchers later ran an entire imaging comparison in software, generating a population of synthetic breasts, simulating both modalities and reading the images with model observers; the result agreed with the direction of the patient trials.5 Physiologically based pharmacokinetic models are routinely submitted to regulators to support dosing in groups never studied directly and to justify not running certain drug-interaction studies; related compartmental models are used to predict where a Targeted drug delivery carrier ends up.
Elsewhere, field models estimate the tissue volume a Deep brain stimulation lead activates in a specific patient's anatomy and are used to guide programming; musculoskeletal simulation scaled to a subject's dimensions and driven by their recorded motion informs orthopaedic surgery and Powered exoskeletons design;6 flow models support the evaluation of circulatory devices including the Artificial heart. Patient-derived Organoids fill the same niche in wet biology, with the advantage of containing the person's actual cells rather than an estimate of their parameters.
Regulators have moved further than clinicians. The FDA published guidance in 2016 on how computational modelling studies should be reported in device submissions,7 and recognises a consensus standard, ASME V&V 40, that scales the credibility a model must demonstrate to the risk of the decision it informs.8 The organising idea is context of use: a model is never validated in general, only for a stated question at a stated level of influence. A simulation used to screen candidate designs and a simulation used to replace a clinical trial are, under that framework, different objects with different burdens.
Law has moved the same way. The FDA Modernization Act 2.0, enacted at the end of 2022, removed the blanket requirement for animal testing of new drugs; in 2025 the agency announced a plan to reduce animal testing for certain classes of biologics, again naming computational methods among the alternatives. Neither change validates any particular model. Both raise the value of building one. Judged on Technology readiness level, the regulatory pathway is more mature than the science travelling down it.
The binding constraint is not compute. It is that a body is not a mechanism with known parameters, and the parameters that would individualise a model are mostly not measurable in a living person.
Identifiability. A model with more free parameters than independent measurements has many parameter sets that fit the data equally well and disagree about the intervention. Fixing the surplus at population values does not remove the problem; it hides it, and turns a personalised model into a population model wearing one patient's anatomy.
No counterfactual to check against. If the model recommends ablating one site and the clinician ablates it, the outcome of the alternative is never observed. Validating a predictive model of an individual therefore requires randomising against ordinary care, which is expensive and rarely done. A related shortage of checkable ground truth slows the validation of Aging biomarkers.
Scope. A model represents what its builders chose to represent. Cardiac mechanics and drug distribution are tractable because the physics is well characterised and the geometry does much of the work. Immune response, metabolism, wound healing and the brain are not in that position, and a whole-body twin needs all of them at once, coupled.
Drift. A person is not a fixed system. Anatomy remodels, drugs change physiology, disease progresses. A model validated at one moment silently expires, and there is usually no signal telling anyone when.
The twin nobody has builtNo published work has produced a model of a whole human body, personalised to an individual, that predicts responses across organ systems. Roadmaps that show one arriving within a decade or two extrapolate from organ-scale successes, and those successes are concentrated exactly where the physics is simple and the anatomy carries the prediction. Cardiology and pharmacokinetics are not a representative sample of physiology.
Mechanism or learningOne camp argues that only mechanistic models can answer interventional questions, because a model fitted to observed care cannot say what would happen under care that was never given. The other argues that mechanistic models will always be underdetermined in an individual, and that large learned models trained across millions of records will predict better in practice even if they explain nothing. Hybrids are common in the literature and are not yet common in anything a regulator has cleared.
A model dense enough to be useful is also a compact encoding of a person's health state, and one that can be queried for conditions the subject has not been told about. Existing protections were written for records and samples, not for parameterised models derived from them; the discrimination problem is the one already mapped by Genetic discrimination, with the added wrinkle that a twin can be re-run against future questions. Models of neural tissue raise the concerns collected under Mental privacy.
Overtrust is the nearer hazard. A simulation produces a specific number with a rendered picture attached, which is a persuasive combination regardless of the model's credibility grade, and clinicians are not generally equipped to audit the assumptions behind it. Liability is unsettled: a wrong prediction from software the clinician cannot inspect sits awkwardly between device failure and clinical judgment. Cost and data availability point the same way as everything else in Access and inequality, since the models are built from imaging and continuous monitoring that many health systems do not provide.
The direction with the most support behind it is not personal twins at all but virtual patient populations: cohorts of synthetic individuals, sampled to span real anatomical and physiological variation, used to test devices and dosing regimens across conditions no trial could enrol. That work has regulatory precedent, a credibility standard, and results. Proponents of the personal version argue it follows once enough individual data is routinely collected, and roadmaps published by research consortia describe a staged path from organ models to coupled ones. Sceptics point out that the missing ingredient is not integration but individual measurement, and that no amount of coupling fixes parameters nobody can observe.
Specific applications keep the idea alive. Long-duration missions are one, where a crew without a physician might carry a model of each member; Space medicine research treats this as attractive and unproven, and an astronaut's physiology changes in ways that are still being characterised, which is precisely the regime where personalisation fails. Learned components are another, with methods from AI drug discovery and structure prediction of the kind used in AI protein design proposed as a source of the parameters mechanistic models cannot measure; the proposal runs ahead of the evidence, since structure prediction returns coordinates rather than the binding affinities and conformational changes a physiological model would need. The benchmark that would settle the field's claims is a randomised comparison in which a personalised model predicts an individual's response to a therapy better than a clinician working from standard information. Trials of that design are rare, and none has been reported for a model spanning more than one organ system. Until that changes, the phrase describes an ambition and a set of organ models, not a technology anyone possesses.
paperGlaessgen, E. and Stargel, D. "The Digital Twin Paradigm for Future NASA and U.S. Air Force Vehicles." 53rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, 2012.↩The paper is about airframes; its framing of a twin as continuously updated from sensor data is the sense medicine borrowed and rarely implements.
paperNørgaard, B.L. et al. "Diagnostic Performance of Noninvasive Fractional Flow Reserve Derived from Coronary Computed Tomography Angiography in Suspected Coronary Artery Disease." Journal of the American College of Cardiology, 2014.↩Diagnostic accuracy against invasive pressure measurement; the study does not show that acting on the model improves outcomes.
paperPrakosa, A. et al. "Personalized virtual-heart technology for guiding the ablation of infarct-related ventricular tachycardia." Nature Biomedical Engineering, 2018.↩Simulated targets were compared with the sites clinicians had actually ablated, in a small retrospective cohort.
paperKovatchev, B.P. et al. "In Silico Preclinical Trials: A Proof of Concept in Closed-Loop Control of Type 1 Diabetes." Journal of Diabetes Science and Technology, 2009.↩The simulator replaced animal testing only for control-algorithm evaluation; the devices themselves still required human trials.
paperBadano, A. et al. "Evaluation of Digital Breast Tomosynthesis as Replacement of Full-Field Digital Mammography Using an In Silico Imaging Trial." JAMA Network Open, 2018.↩Synthetic anatomy sampled from population statistics, not models of individual patients, and a comparison of imaging methods rather than of treatments.
paperDelp, S.L. et al. "OpenSim: Open-Source Software to Create and Analyze Dynamic Simulations of Movement." IEEE Transactions on Biomedical Engineering, 2007. ↩
regulatorU.S. Food and Drug Administration. Reporting of Computational Modeling Studies in Medical Device Submissions: Guidance for Industry and Food and Drug Administration Staff, 2016. ↩
reportAmerican Society of Mechanical Engineers. V&V 40-2018: Assessing Credibility of Computational Modeling Through Verification and Validation: Application to Medical Devices, 2018.↩Credibility is graded against the decision a model informs, so the same model can be adequate for one question and inadequate for another.